XOR Cluster Dataset Overview

• XOR Cluster Dataset is a synthetic binary-classification construction that assigns labels via an exclusive-or rule, creating alternating, nonlinearly separable class regions.
• It encompasses variants from clean point configurations to high-dimensional Gaussian mixtures and noisy clusters, each probing nonlinear decision boundaries and clustering phenomena.
• This benchmark is used to test neural network expressivity and quantum classifier performance under controlled noise models and feature learning scenarios.

The XOR cluster dataset is a family of synthetic binary-classification constructions in which labels are assigned by an exclusive-or rule, so that the two classes occupy alternating regions rather than being globally linearly separable. In the cited literature, this family includes the canonical four-point XOR configuration in $\mathbb R^2$, noisy clustered variants formed by replacing the four corners with Gaussian blobs, high-dimensional Gaussian mixture models with class-conditional centers at $\{\pm \mu_1,\pm \mu_2\}$, and zero-margin Gaussian formulations in which the label is $y=-\operatorname{sgn}(x_1x_2)$. Across these variants, the shared mathematical role of the dataset is to isolate nonlinear expressivity, feature learning, and clustering phenomena that cannot be captured by a single affine decision boundary (Seilkhan et al., 27 Feb 2026, Xu et al., 2023, Braun et al., 30 Jan 2026).

1. Canonical constructions

Several related data-generating mechanisms appear under the XOR-cluster rubric. They differ in ambient dimension, noise model, and whether the cluster structure is explicit in input space or latent in a low-dimensional projection, but they all preserve the alternating XOR label geometry.

Variant	Sampling rule	Label rule
Clean XOR (Dataset A)	Four points at $(0,0)$, $(0,1)$, $(1,0)$, $(1,1)$	$(0,0)\mapsto 0$, $(0,1)\mapsto 1$, $(1,0)\mapsto 1$, $\{\pm \mu_1,\pm \mu_2\}$0
Noisy clustered XOR (Dataset B)	Four clusters around the same corner points with additive Gaussian noise $\{\pm \mu_1,\pm \mu_2\}$1	Each cluster inherits its corner’s XOR label
Continuous XOR (Dataset C)	Points sampled uniformly from $\{\pm \mu_1,\pm \mu_2\}$2	Class $\{\pm \mu_1,\pm \mu_2\}$3 if exactly one coordinate exceeds $\{\pm \mu_1,\pm \mu_2\}$4; main setup uses $\{\pm \mu_1,\pm \mu_2\}$5
Gaussian mixture XOR cluster data	$\{\pm \mu_1,\pm \mu_2\}$6 sampled from $\{\pm \mu_1,\pm \mu_2\}$7 with $\{\pm \mu_1,\pm \mu_2\}$8 and $\{\pm \mu_1,\pm \mu_2\}$9	Positive class from $y=-\operatorname{sgn}(x_1x_2)$0, negative class from $y=-\operatorname{sgn}(x_1x_2)$1, then $y=-\operatorname{sgn}(x_1x_2)$2-corrupted labels
Zero-margin Gaussian XOR	$y=-\operatorname{sgn}(x_1x_2)$3	$y=-\operatorname{sgn}(x_1x_2)$4

The noisy clustered benchmark used for model comparison in two dimensions replaces each corner of the clean XOR dataset with a Gaussian cluster and varies both the noise scale and the sample count per cluster. The experimental sweep uses $y=-\operatorname{sgn}(x_1x_2)$5 and $y=-\operatorname{sgn}(x_1x_2)$6. In effect, the data are four Gaussian blobs centered at the XOR corners, with class labels inherited from the corresponding corner (Seilkhan et al., 27 Feb 2026).

The high-dimensional mixture construction formalizes XOR cluster data as a distribution rather than as a finite geometric toy example. The clean labeled distribution $y=-\operatorname{sgn}(x_1x_2)$7 is generated by first sampling $y=-\operatorname{sgn}(x_1x_2)$8, then drawing $y=-\operatorname{sgn}(x_1x_2)$9 from $(0,0)$0 when $(0,0)$1 and from $(0,0)$2 when $(0,0)$3, with $(0,0)$4. The observed label is then flipped with probability $(0,0)$5, producing a constant-fraction noisy-label regime (Xu et al., 2023).

The zero-margin Gaussian variant removes explicit input-space clusters but retains an XOR partition in the first two coordinates. Writing $(0,0)$6 with $(0,0)$7 and $(0,0)$8, the label depends only on the quadrant of $(0,0)$9. This creates what the paper describes as a low-dimensional XOR cluster structure inside a high-dimensional Gaussian cloud (Braun et al., 30 Jan 2026).

2. Geometry, separability, and boundary structure

The defining geometric property of XOR cluster data is linear inseparability. For the four-point XOR benchmark, a linear classifier of the form

$(0,1)$0

would have to satisfy

$(0,1)$1

$(0,1)$2

These inequalities are contradictory, which proves that XOR cannot be solved by a linear classifier (Seilkhan et al., 27 Feb 2026).

In the Gaussian mixture version, the same obstruction appears at the distributional level: the positive class occupies the opposing clusters at $(0,1)$3, while the negative class occupies the opposing clusters at $(0,1)$4. The paper states explicitly that this is “XOR” because a linear separator cannot separate the classes globally: each class appears in two opposing clusters (Xu et al., 2023).

The zero-margin Gaussian formulation sharpens the boundary geometry further. With

$(0,1)$5

the target boundary in the $(0,1)$6-plane is the union of the coordinate axes. Because the Gaussian distribution places nontrivial mass arbitrarily close to the axes, standard positive-margin arguments fail. The dataset is therefore intrinsically different from linearly inseparable problems that nevertheless admit a positive separation margin (Braun et al., 30 Jan 2026).

This geometry yields two complementary interpretations. In low-dimensional clustered benchmarks, the XOR structure is visible as four alternating blobs. In the zero-margin Gaussian setting, the cluster structure is latent: the distribution is rotationally symmetric in the ambient space, but the label depends only on the first two coordinates and is determined by the quadrant. A plausible implication is that “cluster” in this literature refers not only to visible blob structure in input space, but also to the symmetry classes induced by the XOR labeling rule.

3. Noise models, ambiguity, and benchmark protocol

Noise enters XOR cluster datasets in two distinct ways. The first is feature noise. In noisy clustered XOR, each corner point is replaced by a Gaussian cluster with additive Gaussian noise controlled by $(0,1)$7. Increasing $(0,1)$8 increases overlap and makes the task less cleanly separable, while increasing the number of samples per cluster changes the amount of data available but not the underlying class geometry (Seilkhan et al., 27 Feb 2026).

The second is label noise. In the high-dimensional Gaussian mixture model, the observed distribution is the $(0,1)$9-corrupted version of the clean XOR-cluster distribution: $(1,0)$0 The theory treats $(1,0)$1 as a fixed constant bounded away from $(1,0)$2, which produces a regime in which a model can interpolate noisy training labels while still having to recover the latent cluster structure to generalize (Xu et al., 2023).

In the zero-margin Gaussian setting, ambiguity is geometric rather than label-corruption based. The difficult region is the near-boundary set

$(1,0)$3

This is exactly the region where the projected point $(1,0)$4 is close to one of the axes. The paper shows that for $(1,0)$5,

$(1,0)$6

and interprets generalization in average-case terms: points away from the axes form reliable regions, while points near the axes constitute a persistent error strip (Braun et al., 30 Jan 2026).

The two-dimensional benchmarking protocol is explicit. For Dataset B, the data split is fixed at 80/20 train/test with data seed 42. Model training is repeated with seeds $(1,0)$7, and for some seed-sensitivity analyses an extended range $(1,0)$8 is used. Reported results are typically mean $(1,0)$9 standard deviation across runs. The principal metrics are accuracy and binary cross-entropy (BCE) (Seilkhan et al., 27 Feb 2026).

A common misconception is that XOR benchmarks are exhausted by accuracy scores. The hardware study of the variational quantum classifier shows otherwise: accuracy can remain perfect even when the decision function is measurably distorted. The paper therefore argues that function-level diagnostics matter, especially on hardware (Seilkhan et al., 27 Feb 2026).

4. Feature learning, neuron clustering, and optimization phenomena

XOR cluster data has become a standard setting for analyzing feature learning precisely because linear shortcuts are insufficient. In the noisy-label Gaussian mixture model, a two-layer ReLU network

$(1,1)$0

is trained by gradient descent on empirical logistic loss, with fixed second-layer weights $(1,1)$1 and updates applied only to the first layer. Under the paper’s scaling regime, including

$(1,1)$2

the network first memorizes and only later learns the true cluster features (Xu et al., 2023).

The initial phase is catastrophic overfitting. After the first GD step, the network interpolates the noisy training set,

$(1,1)$3

but on the clean distribution its test error is essentially random: $(1,1)$4 The paper explains this by showing that after one step the network behaves approximately like a linear classifier,

$(1,1)$5

which can fit nearly orthogonal training points but cannot solve the XOR structure (Xu et al., 2023).

The later phase is grokking or benign overfitting. For

$(1,1)$6

the same network still fits all training points but now attains near-zero clean test error: $(1,1)$7 The stated mechanism is feature learning: positive second-layer neurons align with $(1,1)$8, and negative second-layer neurons align with $(1,1)$9 (Xu et al., 2023).

The zero-margin Gaussian analysis refines this picture by introducing neuron block dynamics. A two-layer ReLU network

$(0,0)\mapsto 0$0

trained by SGD on logistic loss,

$(0,0)\mapsto 0$1

develops four coherent neuron blocks aligned with the diagonal directions

$(0,0)\mapsto 0$2

namely $(0,0)\mapsto 0$3 and $(0,0)\mapsto 0$4 (Braun et al., 30 Jan 2026).

The paper defines the initial blocks

$(0,0)\mapsto 0$5

$(0,0)\mapsto 0$6

and studies block masses

$(0,0)\mapsto 0$7

The central claim is that neurons self-organize into four nearly balanced clusters, and that block-level signals evolve coherently even when individual neuron trajectories are highly variable (Braun et al., 30 Jan 2026).

Training is divided into two phases. In Phase I, when outputs are small, the loss is linearized and the signal component grows approximately multiplicatively; a representative law is

$(0,0)\mapsto 0$8

By the end of Phase I, neurons have aligned with one of $(0,0)\mapsto 0$9, and the block imbalance remains small, with

$(0,1)\mapsto 1$0

In Phase II, the block mass evolves approximately as

$(0,1)\mapsto 1$1

where the average margin statistic satisfies

$(0,1)\mapsto 1$2

The theoretical picture is therefore not merely that the network learns a nonlinear separator, but that it learns a four-block internal representation matched to XOR symmetry (Braun et al., 30 Jan 2026).

5. Comparative benchmark results in classical and quantum models

The noisy clustered XOR benchmark has been used to compare linear, classical nonlinear, and variational quantum classifiers under a common protocol. The three model families are logistic regression (LR), a one-hidden-layer MLP with main comparison width $(0,1)\mapsto 1$3, and a two-qubit variational quantum classifier (VQC) with circuit depth $(0,1)\mapsto 1$4. For the VQC, the features are angle-encoded as

$(0,1)\mapsto 1$5

and the ansatz depth is

$(0,1)\mapsto 1$6

with $(0,1)\mapsto 1$7 trainable parameters (Seilkhan et al., 27 Feb 2026).

The benchmark results are centered on the representative setting $(0,1)\mapsto 1$8. Logistic regression reaches train accuracy $(0,1)\mapsto 1$9, test accuracy $(1,0)\mapsto 1$0, and test BCE $(1,0)\mapsto 1$1. The one-hidden-layer MLP achieves train accuracy $(1,0)\mapsto 1$2, test accuracy $(1,0)\mapsto 1$3, and test BCE $(1,0)\mapsto 1$4. The depth-$(1,0)\mapsto 1$5 VQC remains close to chance, with train accuracy about $(1,0)\mapsto 1$6, test accuracy about $(1,0)\mapsto 1$7, and test BCE about $(1,0)\mapsto 1$8. The depth-$(1,0)\mapsto 1$9 VQC reaches train accuracy $\{\pm \mu_1,\pm \mu_2\}$00, test accuracy $\{\pm \mu_1,\pm \mu_2\}$01, and test BCE $\{\pm \mu_1,\pm \mu_2\}$02 analytically or $\{\pm \mu_1,\pm \mu_2\}$03 with 1024 shots (Seilkhan et al., 27 Feb 2026).

These results support the paper’s qualitative conclusion that the XOR task is not about “quantum vs classical” per se, but about expressivity and topology. Logistic regression and the depth-$\{\pm \mu_1,\pm \mu_2\}$04 VQC fail because they cannot represent the required nonlinear separation reliably, whereas the MLP and the depth-$\{\pm \mu_1,\pm \mu_2\}$05 VQC succeed because they can represent nonlinear decision boundaries (Seilkhan et al., 27 Feb 2026).

Robustness analyses reinforce the same point. As $\{\pm \mu_1,\pm \mu_2\}$06 increases, LR remains low and nearly constant, VQC $\{\pm \mu_1,\pm \mu_2\}$07 degrades smoothly but remains weak, and MLP $\{\pm \mu_1,\pm \mu_2\}$08 together with VQC $\{\pm \mu_1,\pm \mu_2\}$09 attain the best performance at low noise and degrade gradually as noise rises. As the number of samples per cluster increases, LR and VQC $\{\pm \mu_1,\pm \mu_2\}$10 stay relatively low, whereas MLP $\{\pm \mu_1,\pm \mu_2\}$11 and VQC $\{\pm \mu_1,\pm \mu_2\}$12 remain at $\{\pm \mu_1,\pm \mu_2\}$13 accuracy across all tested sizes. Seed sensitivity is small overall, but BCE consistently favors the MLP over the depth-$\{\pm \mu_1,\pm \mu_2\}$14 VQC even when both have perfect accuracy (Seilkhan et al., 27 Feb 2026).

The decision-boundary visualizations are consistent with the numerical metrics. LR yields a linear boundary and cannot carve out the XOR pattern; MLP $\{\pm \mu_1,\pm \mu_2\}$15 learns a nonlinear boundary that cleanly separates the four XOR regions; VQC $\{\pm \mu_1,\pm \mu_2\}$16 produces only a coarse nonlinear partition; and VQC $\{\pm \mu_1,\pm \mu_2\}$17 yields a clearly nonlinear surface that matches the XOR structure (Seilkhan et al., 27 Feb 2026).

The hardware execution study isolates another feature of the dataset as a benchmark. A trained depth-$\{\pm \mu_1,\pm \mu_2\}$18 VQC is run on an IBM Quantum backend using the clean XOR dataset, with hardware inference only, a $\{\pm \mu_1,\pm \mu_2\}$19 grid, and 1024 shots per point. The global XOR pattern is preserved, but the hardware decision surface shows structured local distortions compared with the ideal simulator. The paper quantifies the discrepancy by a mean absolute deviation of about $\{\pm \mu_1,\pm \mu_2\}$20 between hardware and finite-shot simulation in a representative run. This suggests that XOR cluster datasets can probe not only classification accuracy but also the fidelity of learned decision functions (Seilkhan et al., 27 Feb 2026).

6. Relation to clustering in random XOR equations

A distinct XOR literature uses the language of clustering in a combinatorial rather than supervised-learning sense. There, the object is a random system of $\{\pm \mu_1,\pm \mu_2\}$21 linear equations over $\{\pm \mu_1,\pm \mu_2\}$22 variables in $\{\pm \mu_1,\pm \mu_2\}$23, equivalently random $\{\pm \mu_1,\pm \mu_2\}$24-XORSAT, and “clusters” are connected components of the solution space under bounded Hamming moves rather than groups of input samples (Gao et al., 2013, Gao et al., 2015).

The random model $\{\pm \mu_1,\pm \mu_2\}$25 has $\{\pm \mu_1,\pm \mu_2\}$26 Boolean variables, each possible $\{\pm \mu_1,\pm \mu_2\}$27-tuple chosen independently with probability

$\{\pm \mu_1,\pm \mu_2\}$28

and a uniformly random right-hand side in $\{\pm \mu_1,\pm \mu_2\}$29. The underlying hypergraph is $\{\pm \mu_1,\pm \mu_2\}$30-uniform, with variables as vertices and equations as hyperedges. For $\{\pm \mu_1,\pm \mu_2\}$31, the clustering threshold is the $\{\pm \mu_1,\pm \mu_2\}$32-core threshold

$\{\pm \mu_1,\pm \mu_2\}$33

with

$\{\pm \mu_1,\pm \mu_2\}$34

in one formulation, and

$\{\pm \mu_1,\pm \mu_2\}$35

in the alternative formulation used for $\{\pm \mu_1,\pm \mu_2\}$36 (Gao et al., 2013, Gao et al., 2015).

Above the threshold by a constant amount, prior work showed that the satisfying assignments split into well-connected, well-separated clusters: within a cluster one can move between any two solutions by changing only $\{\pm \mu_1,\pm \mu_2\}$37 variables per step, while moving between clusters requires a linear-size step, $\{\pm \mu_1,\pm \mu_2\}$38 (Gao et al., 2013). The cluster description is hypergraph-based. A cluster is determined by a solution on the $\{\pm \mu_1,\pm \mu_2\}$39-core together with satisfying extensions outside the $\{\pm \mu_1,\pm \mu_2\}$40-core, modulo the small ambiguity generated by core flippable cycles. In the formal description, the solution clusters are the cycle-equivalence classes (Gao et al., 2013, Gao et al., 2015).

The near-threshold regime changes the internal geometry. For

$\{\pm \mu_1,\pm \mu_2\}$41

the cluster connectivity parameter becomes

$\{\pm \mu_1,\pm \mu_2\}$42

meaning that any two solutions in the same cluster can be connected by a sequence of satisfying assignments differing on at most $\{\pm \mu_1,\pm \mu_2\}$43 variables at each step. The upper bound is tight up to the constant in the exponent: the paper gives a lower obstruction of the form “not $\{\pm \mu_1,\pm \mu_2\}$44-connected” in the chosen regime. Distinct clusters remain strongly separated, and any path between them must contain a step of size at least

$\{\pm \mu_1,\pm \mu_2\}$45

Below the threshold, for

$\{\pm \mu_1,\pm \mu_2\}$46

all solutions remain in one cluster and the whole solution space is $\{\pm \mu_1,\pm \mu_2\}$47-connected (Gao et al., 2013, Gao et al., 2015).

The mechanism is the parallel $\{\pm \mu_1,\pm \mu_2\}$48-stripping process on the underlying hypergraph. Repeatedly removing all vertices of degree $\{\pm \mu_1,\pm \mu_2\}$49 until only the $\{\pm \mu_1,\pm \mu_2\}$50-core remains yields a stripping number and maximum stripping depth of

$\{\pm \mu_1,\pm \mu_2\}$51

when $\{\pm \mu_1,\pm \mu_2\}$52, instead of $\{\pm \mu_1,\pm \mu_2\}$53 at constant supercritical density. The longer stripping chains mean that changing a single free variable can propagate through a polynomial-size dependency structure. In the XORSAT setting, this produces polynomial-scale coordinated changes within a cluster (Gao et al., 2013).

This usage of “XOR clusters” is mathematically separate from the supervised XOR-cluster datasets used in machine learning, but the two literatures share a structural theme: in both cases, XOR induces nontrivial geometry that is invisible to purely linear descriptions. In one case the geometry concerns class regions in input space; in the other it concerns connectivity and separation in the space of satisfying assignments.